sdk: compiled messages deeper than a list in a variable keep the variables of each turn #113

In the fix of #97 (#112), compile() looks for compiled messages in its variables: a message as a variable value, or as an item of a list (Python: a list or tuple) in a variable. When it finds one, the compiles two steps back remove their variables. Thus a chat history of compiled messages keeps the variables of its last two turns. In a test of 2000 turns, 2 turns kept their variables.

compile() does not look deeper in a variable. Examples: Python prompt.compile(context={"history": history}), or a TypeScript array of objects that hold compiled messages. In these cases each compile keeps its nested history while its messages live, and memory increases with the square of the turns.

Options:

1. Walk lists, tuples, and plain dicts or objects (exact types only, so no app code runs) to a fixed depth, with a cap on the count of nodes.
2. Put a hard cap on the count of compiles that keep variables. Risk: on a busy server, a long agent loop that compiles one time loses its variables.

Acceptance: a gc test in TypeScript (v8.setFlagsFromString + WeakRef) and in Python (gc + weakref) with the history in a nested dict or object shows the same count of kept turn variables after 150 and after 300 turns. The existing td.prompt.variables tests continue to pass.